Seminar Series: Inverse-design of nonlin. mech. metamaterials via video denoising diffusion models

"Inverse-design of nonlinear mechanical metamaterials via denoising diffusion models"

Abstract:
The accelerated inverse design of nonlinear material characteristics, such as identifying a material with a predetermined stress-strain response over a finite deformation path, offers significant opportunities for addressing key issues in fields like soft robotics, biomedical implants, and energy absorption. Machine learning models have received significant attention and demonstrated success in obtaining these highly nonlinear mappings, though they have largely been restricted to linear properties, like the directional Young’s modulus, and seldom to more advanced settings as the ones mentioned above.
We here address this challenge and present a novel framework that leverages recent advances in generative models. Specifically, we develop a diffusion model trained on the full-field video data of periodic stochastic structures obtained via finite-element (FE) simulations to predict the deformation and stress response of these structures in the finite-strain regime. By conditioning the model on a target stress-strain response, it generates a large variety of structures that closely match the desired curve for compressive strains of up to 20%. Unlike commonly encountered black-box models, our framework intrinsically provides a complete full-field estimate of the expected deformation path, including the internal stresses which closely align with results obtained via simulations. This eliminates the need for a complex setup of FE frameworks to verify the designed structures and their stress-strain responses. This work has thus the potential to greatly simplify and accelerate the identification of materials with complex target properties.

Presented by Jan-Hendrik Bastek as part of the USACM Student Chapter Seminar Series on 03 November 2023

Bio:
Jan-Hendrik Bastek studied Industrial Engineering at the Technical University of Braunschweig (Germany) and Mechanical Engineering at ETH Zürich (Switzerland), receiving his M.Sc. degree in 2019. His master thesis, which was awarded the ETH medal, introduced a new theoretical-numerical framework to predict the behavior of viscoelastic beam structures. After a brief stint in the consulting industry, he joined the Mechanics & Materials Lab at ETH Zürich as a Doctoral Student under the supervision of Prof. Dennis Kochmann in 2020 and currently spends a semester as a Visiting Scholar in Prof. Steve Sun’s group at Columbia University.
His research interests lie at the intersection of mechanics and scientific machine learning, where he is investigating both data-driven (e.g., inverse design of metamaterials) and data-free methods such as Physics-Informed Neural Networks (e.g., on non-Euclidean domains for mechanical shell theory) as alternatives to established numerical frameworks.